Leave a working agent on a remote box.
Dispatch an autonomous agent to a VM, let it work while you're away, and attach back any time to inspect and steer it. The agent replicates itself onto the box — it's one static binary, no Python, no runtime, nothing to install on the target. Built in machin (MFL).
# get a key for your provider (Anthropic, or any OpenAI-compatible endpoint) export ROAM_API_KEY=sk-... # send an agent to a box and walk away roam send --to my-vm --allow-shell --confirm --verify \ --goal "clone the repo, run its tests, write results.md" # watch it think and act, live over ssh roam attach my-vm # it parks on anything destructive — approve out-of-band roam approve my-vm roam stop my-vm
A genuine tool-use loop — read, write, run shell — until the goal is met. Claude or any OpenAI-compatible model (OpenRouter, …).
Hard token & iteration budgets, a workdir sandbox, shell off by default, and a full journal of every step.
The confirm-gate parks destructive commands for async approval; a deny budget stops a stubborn agent. Same interface for humans and supervising agents.
An optional judge checks the goal against real evidence — the working directory and action log — not the agent's word. Fail-open, so it never blocks correct work.
The hosted hub at roam.intrane.fr handles API keys, billing, and worker coordination for you. Workers hold no model keys — every LLM call goes through the hub's metered proxy. Pay per run (€0.01 + token cost × 1.3), no monthly fee. Route agents to specific workers with labels. Join the waitlist for early access.
Get early access →